bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

1 TITLE:

2 Inferring the role of habitat dynamics in driving diversification: evidence for a species pump

3 in Lake Tanganyika cichlids

4 AUTHORS: Thijs Janzen1,2,3*, Rampal S. Etienne2

5

6 AUTHOR AFFILIATION

7 1. Max Planck Institute for Evolutionary Biology, Department of Evolutionary Theory,

8 August-Thienemann-Straße 2, 24306, Plön, Germany.

9 2. University of Groningen, Groningen Institute for Evolutionary Life Sciences, Box

10 11103, 9700 CC Groningen, The Netherlands

11 3. Carl von Ossietzky University of Oldenburg, Institute for Biology, Department of

12 Molecular Ecology, Carl-von-Ossietzky Straße 9, 26111, Oldenburg, Germany

13

14 4. *Corresponding author: Thijs Janzen, Carl von Ossietzky University of Oldenburg,

15 Institute for Biology, Department of Molecular Ecology, Carl-von-Ossietzky Straße 9,

16 26111, Oldenburg, Germany

17 e-mail: [email protected]

18 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

19 ABSTRACT

20 Geographic isolation that drives speciation is often assumed to slowly increase over time, for

21 instance through the formation of rivers, the formation of mountains or the movement of

22 tectonic plates. Cyclic changes in connectivity between areas may occur with the

23 advancement and retraction of glaciers, with water level fluctuations in seas between islands

24 or in lakes that have an uneven bathymetry. These habitat dynamics may act as a driver of

25 allopatric speciation and propel local diversity. Here we present a parsimonious model of the

26 interaction between cyclical (but not necessarily periodic) changes in the environment and

27 speciation, and provide an ABC-SMC method to infer the rates of allopatric and sympatric

28 speciation from a phylogenetic tree. We apply our approach to the posterior sample of an

29 updated phylogeny of the Lamprologini, a tribe of cichlid fish from Lake Tanganyika where

30 such cyclic changes in water level have occurred. We find that water level changes play a

31 crucial role in driving diversity in Lake Tanganyika. We note that if we apply our analysis to

32 the Most Credible Consensus (MCC) tree, we do not find evidence for water level changes

33 influencing diversity in the Lamprologini, suggesting that the MCC tree is a misleading

34 representation of the true species tree. Furthermore, we note that the signature of habitat

35 dynamics is found in the posterior sample despite the fact that this sample was constructed

36 using a species tree prior that ignores habitat dynamics. However, in other cases this species

37 tree prior might erase this signature. Hence we argue that in order to improve inference of the

38 effect of habitat dynamics on biodiversity, phylogenetic reconstruction methods should

39 include tree priors that explicitly take into account such dynamics.

40

41 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

42 INTRODUCTION

43

44 Environmental changes such as the formation of mountain ridges, the formation of rivers and

45 the movement of tectonic plates have long been known to be important drivers of speciation

46 (Coyne and Orr 2004). Repeated environmental changes may thus lead to diversification

47 patterns. Cyclic changes in the environment can cause populations to continuously switch

48 between an allopatric and sympatric stage, providing a continuously renewed potential for

49 speciation. And these cyclic changes can in turn drive diversity towards levels unexpected

50 given the current geography, sometimes referred to as a “species pump” (Heaney 1985;

51 Rossiter 1995). Examples of species pumps include environmental fluctuations fragmenting

52 habitats on the slopes of mountains (Weir 2006; Sedano and Burns 2010; Hutter et al. 2013),

53 glaciations and postglacial secondary contacts (Barnosky 2005), sea level changes causing

54 the fusion and fragmentation of islands (Glor et al. 2004; Thorpe et al. 2008, but see

55 Papadopoulou and Knowles 2015), and fluctuations in water level causing fragmentation and

56 fusion of lakes with uneven bathymetry, as in the African Rift Lakes (Cohen et al. 1997b;

57 Alin et al. 1999; McGlue et al. 2008; Ivory et al. 2016).

58

59 The African Rift Lakes provide a good starting point in studying the interplay between cyclic

60 habitat dynamics and speciation, because they have been subject to frequent water level

61 changes (Cohen et al. 1997b; Alin and Cohen 2003; Ivory et al. 2016), and are well known

62 for their tremendous biodiversity (Seehausen 2000, 2006; Turner et al. 2001; Wagner et al.

63 2012, 2014; Brawand et al. 2014). An estimated number of 2000 cichlid fish species (Turner

64 et al. 2001) have evolved in the African Rift Lakes over the past 10 million years (Genner et

65 al. 2007; Meyer et al. 2016), and comprise one of the most spectacular adaptive radiations bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

66 (Seehausen 2006). The most prominent water level changes took place in Lake Tanganyika,

67 where the water level has dropped substantially on multiple occasions over the past million

68 years, sometimes splitting the lake into multiple smaller lakes (Lezzar et al. 1996; Cohen et

69 al. 1997a, 2007). Being the oldest lake of the three large rift lakes (Cohen et al. 1993), Lake

70 Tanganyika contains the highest behavioral diversity (Konings 2007) and is the only lake

71 with a highly resolved phylogeny for cichlid fish. Evidence for the influence of changing

72 water levels comes from analysis of mitochondrial DNA, which shows that for Tropheus

73 species, some populations have experienced secondary contact upon changes in water level,

74 potentially increasing genetic diversity and driving speciation (Sturmbauer et al. 2001;

75 Koblmüller et al. 2011; Sefc et al. 2017). Similar patterns were found for Variabilichromis

76 moorii and Ophthalmotilapia nasuta (Sturmbauer et al. 2001), Telmatochromis temporalis

77 (Winkelmann et al. 2016), and Altolamprologus (Koblmüller et al. 2016). Comparison of

78 mitochondrial DNA between populations from deep and shallow areas emphasizes that the

79 deep areas are habitats that are more persistent over time, with lower genetic variation

80 (Nevado et al. 2013). Furthermore, Eretmodus lineages identified using mitochondrial DNA

81 are strongly associated with the bathymetric basins of Lake Tanganyika (Verheyen et al.

82 1996), suggesting that they have independently diversified at low water level.

83

84 Aguilée et al. (2013) developed a model for the African Rift Lakes in which populations at

85 different locations diverge from each other depending on the local habitat, and at the same

86 time allowed for sympatric speciation by implementing assortative mating that allows for a

87 single branching point in trait values. Over time the different locations become separated or

88 are reconnected, and this may drive the formation of new species. The authors conclude that

89 stable numbers of diversity are best obtained by a fragmented habitat with recurrent merged

90 states and rapid fluctuations. However, Aguilée et al. (2013) do not compare their results to bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

91 empirical data. By contrast, Pigot et al (2010) used a spatially explicit model of landscape

92 fragmentation, where consecutive splitting of species’ geographic ranges drives speciation,

93 and compared phylogenies generated with their model, with known bird phylogenies. They

94 found that including this geographical context of speciation explains a large part of the

95 features exhibited by the reconstructed avian trees. Hence, including a geographical context

96 of speciation seems a promising research avenue.

97

98 Here, we provide a method to infer whether or how cyclic changes in the environment

99 influence both the generation and the maintenance of biodiversity. We use an extension of the

100 standard constant-rates birth-death model. Because deriving an expression for the likelihood

101 of this model for a given set of phylogenetic branching times is difficult, but simulation of

102 phylogenies under the model is easy, we used approximate Bayesian computation (ABC)

103 based on sequential Monte Carlo sampling (SMC) to estimate parameters from phylogenies.

104 We applied our approach to an updated phylogeny of the Lamprologini, a tribe of cichlid fish

105 from Lake Tanganyika in order to assess the importance of these habitat dynamics in shaping

106 the current biodiversity of cichlids in Lake Tanganyika.

107

108

109 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

110 METHODS

111 Model

112 To model the interaction between environmental change and speciation, we envisage a lake

113 that consists of a single pocket at high water level, but that splits into two pockets when the

114 water level drops. When the water level drops, we assume that all species distribute

115 themselves equally over the two pockets; similarly, when the water level rises, all species

116 previously contained in the two pockets are combined into the single pocket. Allopatric

117 speciation can only occur when the water level is low. We assume a constant probability rate

118 for allopatric speciation, and hence the waiting time until the next speciation event is

119 exponentially distributed. After this waiting time, one of the two incipient species in either

120 pocket can speciate into a new species. If this allopatric speciation does not occur before the

121 water level rises again, i.e. reflecting that there has not been enough genetic divergence, the

122 two incipient species in the two pockets merge back into one species. This is conceptually

123 similar to the idea of protracted speciation (Etienne and Rosindell 2012): the water level drop

124 initiates the speciation process whereas the allopatric speciation event is the completion of

125 speciation under the protracted speciation model. Sympatric speciation can always occur in

126 our model, either at high water level in the lake, or in both pockets when the water level is

127 low. Extinction is considered to be a background process that occurs locally, i.e. within a

128 pocket. If the water level is high, this causes extinction of a species, if the water level is low,

129 this causes local extinction in one of the pockets.

130 We implemented our model using a Gillespie algorithm, where the time steps are chosen

131 depending on the rate of possible events. In the model there are five possible events:

132 1) A water level change event, inducing incipient species or merger of incipient species.

# 133 2) Sympatric speciation event at high water level, with rate . bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

' 134 3) Sympatric speciation event at low water level, with rate .

' 135 4) Allopatric speciation(-completion) event, at low water level, with rate 

136 5) Extinction event, with rate

137 When the water level drops, all species distribute themselves over both pockets. Thus,

138 immediately after a water level drop, the number of incipient species is equal to twice the

139 number of species. When the water level rises, all incipient species that belong to the same

140 species merge into a single species. During a sympatric speciation event, a single species

141 splits into two new species, and the original (incipient) species is consumed in the process.

142 Here we assume that local disruptive selection causes divergence, similar to the

143 implementation of speciation by Aguilee (Aguilée et al. 2011, 2013). If sympatric speciation

144 occurs when the water level is low, the species in the other pocket is retained, and thus three

145 new lineages arise: the first branching point occurs at the water level drop while the second

146 occurs at the sympatric speciation event (Figure 1).

147

148 Figure 1. Schematic representation of the consequences of the three different types of 149 speciation. Time proceeds from left to right. The dotted blue lines indicate water level 150 changes. (a) During a sympatric speciation at high water level event, diversification is not 151 aligned with any associated water level change. (b) During allopatric speciation at low water 152 level, speciation initiation (incipient species are indicated with a dotted line) coincides with 153 the water level drop, causing branching events (if speciation-completion occurs before water 154 level rise) to line up in time. Branching events are conditionally independent of the time of 155 speciation completion, hence, even when the actual speciation completion events occur at 156 different time points, branching events in the species tree are identical. (c) During a sympatric bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

157 speciation event at low water, the speciation event is independent of the water level changes. 158 Because the original species is consumed in the process, a new branching event is also added 159 at the water level change event. Hence, both speciation-completion (b) and sympatric 160 speciation at low water level (c) cause branching times to line up at the time of water level 161 drop. Please note that (a) and (c) represent the reconstructed species tree, but (b) does not; the 162 reconstructed species tree for (b) would not show the branching event in the uppermost part 163 of the tree. 164 During an extinction event, one (possibly incipient) species is removed from the simulation.

165 If the water level is low, this need not lead to the extinction of a species, because the sister

166 incipient species might remain in the other pocket, ensuring survival of the species.

167 Maximum Likelihood

168 Without water level changes, our model reduces to the constant rates birth-death model (Nee

169 et al. 1994). As a reference therefore, we estimated parameters of the standard birth-death

170 model using Maximum Likelihood. The likelihood of the birth-death model was calculated

171 using the function “bd_ML” from the R package DDD. (Etienne et al. 2012).

172

173 Fitting the model to empirical data

174 We performed two different fitting procedures: firstly, we performed a model selection

175 procedure, where three different water level scenarios were fitted simultaneously to the data

176 (more information about the chosen scenarios can be found in the next section). The model

177 selection procedure simultaneously estimates parameter estimates and assesses the fit of the

178 models. However, because the model selection procedure primarily samples the best fitting

179 model (by design), it does not allow for the comparison of parameter estimates across

180 different models. Therefore, we also fitted the three different water level scenarios

181 independently to the empirical data, and obtained posterior distributions for the parameters

182 relevant to these scenarios. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

183 We fitted our model to 100 trees randomly sampled from the MCMC chain obtained from the

184 *BEAST analysis (see below), and to the Most Credible Consensus (MCC) tree.

185 Water level scenarios

186 The main focus of our approach is to assess the impact of water level changes on the

187 diversification rate. Lake Tanganyika experienced low water level stands 35 - 40 k years ago

188 (kya) (-160 meter), 169 - 193 kya (-250 meter), 262 - 295 kya (-350 meter) , 363 - 393 kya (-

189 350 meter) and 550 - 1100 kya (-650 – 700 meter) (Lezzar et al. 1996; Cohen et al. 1997a).

190 The southern and northern basin of Lake Tanganyika are separated from each other by a ridge

191 at a depth of 500 meter below present level. Although some of these water level changes may

192 not have split up the lake completely, we assume here that these water level changes still

193 caused sufficient disruption of migration between the northern and southern basin, to be

194 equivalent to physical separation. Consequently, high water levels occurred between 0 – 35

195 kya, 40 – 169 kya, 193 – 262 kya, 295 - 363 kya and 393 – 550 kya. Unfortunately the

196 geological record does not reveal whether any low water level stands occurred beyond 1.1

197 million years ago (Ma). This leaves us with two alternative scenarios: either no low water

198 level stands occurred beyond 1.1 Ma, or these low water level stands have not been preserved

199 accurately in the geological record.

200 In order to capture these two scenarios we performed inference using two alternative water

201 level implementations. Firstly we used the exact literature values, assuming a high water

202 level stand until 1.1 Ma. We refer to this scenario as LW (Literature Waterlevels). Secondly

203 we assumed that before 1.1. Ma, water level changes occurred at the same average rate of

204 water level change in the most recent 1.1 million years. In the recent 1.1 million years, the

205 lake experienced 5 high water level stands, and 5 low water level stands, which amounts to

206 10 water level changes in total. To extrapolate water level changes to more than 1.1 Ma, we bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

207 drew waiting times until the next water level change from an exponential distribution with

208 rate 10. We refer to this scenario as EW (Extrapolated Water levels). Thirdly we also tested

209 the null expectation: no effect of water level changes on speciation, we refer to this scenario

210 as NW (No Water levels). Without water level changes, the model reduces to the constant-

211 rates birth death model.

212 Parameter estimation

213 To fit the model to empirical data we used Approximate Bayesian Computation, in

214 combination with a Sequential Monte Carlo scheme (ABC-SMC) (Toni et al. 2009).

215 As summary statistics for the ABC analysis we chose the normalized Lineages Through Time

216 statistic (Janzen et al. 2015), tree size, Phylogenetic Diversity (AvPD, Schweiger et al. 2008)

# ' ' 217 and the γ statistic (Pybus and Harvey 2000). On all parameters (. , . ,  , ) we chose

218 uniform priors U(-3, 2), on a 10log scale, such that the eventual prior distribution spans (10-3,

219 102). A 10log scale was chosen to explore parameter space uniformly, and put extra emphasis

220 on low values. The standard deviation of the normal distribution used to perturb the

221 parameters was chosen to have a mean of 0, and a standard deviation of 0.05 (on the 10log

222 transformed parameter), and we updated one parameter each time (e.g. jumps were only made

223 in one dimension, to avoid extremely low acceptance rates). The number of particles used per

224 SMC step was 10,000, where a particle is a data structure containing the model choice and

225 the parameter estimates. To assess the fit of the model to the data we calculated the Euclidian

226 distance between the summary statistic of the simulated data and the empirical data. To

227 ensure that the differences in summary statistics were on the same scale, we normalized the

228 differences. Differences were normalized by dividing each difference by the standard

229 deviation of that summary statistic of 1,000,000 trees simulated using parameter values

230 sampled from the prior.

231 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

232 Model selection

233 To identify which model best explains the data, we performed ABC model selection, as

234 described in Toni et al. (2009; 2010). The main difference between standard ABC-SMC and

235 ABC-SMC including model selection is that the latter adds one parameter, which keeps track

236 of the model. As jumping kernel between models we assumed a 50% probability of staying at

237 the same model, and a 25% probability of jumping to either other model. We assumed a

238 uniform prior across all three models; this translates to a probability of 1/3 for each model in

239 the first iteration of the ABC-SMC procedure, and hence an expected number of 3333

240 particles assigned to each model in the first iteration. This reversible jump ABC-SMC model

241 selection procedure results in a posterior distribution over the three models, where the model

242 with most support is the model selected most across all particles. We can calculate the Bayes

243 factor by taking the ratio of the number of particles assigned to the respective models (Toni et

244 al. 2009). For example, the Bayes factor of LW/EW is the number of particles assigned to the

245 model with literature water level changes divided by the number of particles assigned to the

246 model with extrapolated water level changes. Because a model can receive zero particles, we

247 set the Bayes factor for each model compared to the model with zero particles to the

248 maximum support possible, which is the total number of particles: 10,000. To calculate the

249 posterior support for a model, we calculate 2 ln (Bayes factor), following Kass and Raftery

250 (1995). A transformed Bayes factor over a value of 2 then corresponds to substantial support

251 for the considered model (Kass and Raftery 1995).

252 Model selection validation

253 To assess whether our ABC-SMC method can accurately infer the correct model, we

254 simulated 100 datasets for each model (NW, LW & EW), with parameter values drawn from

255 the prior. We report the median Bayes factor across the 100 replicates. If our method can bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

256 accurately infer the correct model, we expect the median Bayes Factor (after 2 ln

257 transformation) to be above 2 when comparing posterior support for the model with which

258 the data was simulated to the other two models.

259

260 Measurement uncertainty

261 A phylogeny generated with a high rate of allopatric speciation and a high rate of water level

262 changes tends to have multiple speciation events that are aligned in time (Figure 1, b). This is

263 due to the fact that the onset of speciation is given by the time of water level change.

264 Phylogenetic reconstruction methods such as BEAST (Bouckaert et al. 2014) currently do not

265 allow for simultaneous branching events. Hence, when fitting the model, trees are generated

266 that are by definition dissimilar from the empirical tree constructed using BEAST, even if

267 underlying events are close to the original events. To circumvent this we perturbed the

268 branching time of each node in the trees simulated using our model. In this way speciation

269 events that were previously aligned in time now occur on slightly different time points, as in

270 a tree from a BEAST analysis. We perturbed branching times by adding a random number

271 drawn from a truncated normal distribution with mean 0, standard deviation , truncated by

272 the minimum distance to either the daughter or the parent species. If there were no daughter

273 lineages present, and the node gave rise to an extant species, the normal distribution was

274 truncated to the minimum distance to the parent or the present time. Nodes were perturbed

275 from past to present (leaving the crown in place, to ensure a phylogenetic tree with the same

276 age as the empirical tree). The standard deviation of the perturbation kernel was included as

277 an extra parameter to be inferred, with a uniform prior on (10-3,100).

278 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

279 Empirical data

280 We fitted our model to the phylogenetic tree of the tribe of Lamprologini, the most diverse

281 tribe within Lake Tanganyika, containing 79 species of cichlids in Lake Tanganyika (Day et

282 al. 2007; Koblmüller et al. 2007; Sturmbauer et al. 2010). The Lamprologini are endemic to

283 Lake Tanganyika and its surrounding rivers and all species are substrate brooders with shared

284 paternal and maternal care. In contrast to the mouthbrooding species from the

285 Haplochromini, the Lamprologini show little sexual dimorphism and dichromatism, which

286 are well-known indicators for sexual selection (Kraaijeveld et al. 2011). We therefore expect

287 that the Lamprologini is a good candidate for picking up signals from water level changes.

288 We reconstructed a new Lamprologini tree following the workflow of the most complete

289 Lamprologini tree to date, which is a consensus tree based on the mitochondrial ND2 gene

290 (Sturmbauer et al. 2010), but we added three newly described species (Lepidiolamprologus

291 mimicus (Schelly et al. 2007), Neolamprologus timidus (Kullander et al. 2014b) and

292 Chalinochromis cyanophleps (Kullander et al. 2014a)). Using phyloGenerator (Pearse and

293 Purvis 2013), we downloaded sequences from GenBank for nine genes (GenBank access

294 numbers can be found in the Supplementary Information). Genes were selected on the basis

295 of species coverage (at least 25% of the 79 Lamprologini species for which molecular data is

296 available), and whether or not the gene was crucial for inclusion of a species (e.g. for a

297 number of species, the only available gene was ND2). After selection, our full dataset

298 consisted of three mitochondrial genes: the NADH dehydrogenase subunit 2 (ND2 gene,

299 sequences from Kocher et al. 1995; Klett and Meyer 2002; Clabaut et al. 2005; Duftner et al.

300 2005; Schelly et al. 2006; Day et al. 2007; Koblmüller et al. 2007, 2016; Schwarzer et al.

301 2009; Wagner et al. 2009; Sturmbauer et al. 2010; O’Quin et al. 2010; Kullander et al. 2014b;

302 Weiss et al. 2015). The cytochrome b (cytb) gene (sequences from Salzburger et al. 2002;

303 Nevado et al. 2009; Wagner et al. 2009; O’Quin et al. 2010; Matschiner et al. 2011, 2016; bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

304 Kullander et al. 2014b; Shirai et al. 2014) and the cytochrome c oxidase subunit I (COI gene,

305 sequences from Sparks and Smith 2004; Nevado et al. 2013; Kullander et al. 2014a, 2014b;

306 Breman et al. 2016; Matschiner et al. 2016) and six nuclear genes: the nuclear locus 38A

307 (38A, sequences from Muschick et al. 2012; Meyer et al. 2016), the 18S ribosomal RNA

308 internal-transcribed spacer 1–2 with 5.8S and 28S ribosomal RNA partial sequences (18S,

309 sequences from Nevado et al. 2009; Koblmüller et al. 2016), the recombinase activating

310 protein 1 (rag1, sequences from Clabaut et al. 2005; Nevado et al. 2009; Kullander et al.

311 2014b; Shirai et al. 2014; Koblmüller et al. 2016; Meyer et al. 2016), the endothelin receptor

312 B1 gene (ednrb1, sequences from Muschick et al. 2012; Santos et al. 2014), the ribosomal

313 protein S7 (rps7, sequences from Schelly et al. 2006; Meyer et al. 2016)) gene and the rod

314 opsin gene (RH1, sequences from Sugawara et al. 2002; Spady et al. 2005; Nagai et al. 2011;

315 Meyer et al. 2015). GenBank access numbers for the used sequences can be found in the

316 supplementary material.

317 Sequences were aligned using MAFFT (setting: --auto) (Katoh and Standley 2013), and

318 subsequently, sequences were cleaned using trimAI (sites with >80% data missing were

319 removed, e.g. setting –gt 0.2) (Capella-Gutiérrez et al. 2009). Rather than concatenating the

320 alignments, we partitioned the data into subsets with independent sequence evolution models,

321 which is more suitable for a dataset which is expected to show incomplete lineage sorting or

322 hybridization (Meyer et al. 2016). To prepare alignments for use with partitionFinder,

323 alignments were combined using SequenceMatrix 1.8 (Vaidya et al. 2011). The best

324 partitioning found by partitionFinder 2.1.1 (Lanfear et al. 2012, 2016), partitioned the data

325 into 5 subsets (unlinked branches, AICc selection criterion), with all nuclear genes into one

326 subset (Rps7, ednrb1, 38A, 18S, RAG1 and RH1), with substitution model HKY+I+Γ. The

327 remaining three mitochondrial genes (ND2, COI and cytb) were placed in separate subsets,

328 each with a GTR+I+ Γ substitution model. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

329 Using *BEAST (Heled and Drummond 2010) within the BEAST 2 package (Bouckaert et al.

330 2014), we inferred the time-calibrated species tree. We used an uncorrelated log-normal

331 relaxed clock and applied two calibration points. Firstly, we calibrated the crown of the

332 Lamprologini to be 4 million years old (log-normal prior, mean of 4 Myr 95% conf interval:

333 [3, 5]), based on the results from Meyer et al. (2016). Secondly, we included two riverine

334 Lamprologini species (L. congoensis and L. teugelsi), and calibrated the onset of their

335 branching event at 1.7 Ma (offset 1.1, log normal distribution with mean 1.7, 95% conf

336 interval [1.15, 3.47], “use originate = true”), following Koblmuller (2010). We applied 1/X

337 priors on the clock rates, and log-normal priors on the substitution rates. All other priors were

338 left at their default setting. As tree model we used the birth death model. The used BEAST

339 configuration file (the Beauti xml) can be found in the supplementary material.

340 We ran 10 independent STARBEAST MCMC chains, of 750M trees each. Each chain was

341 verified to have ESS values of at least 100 for all parameters. The first 10M trees were

342 pruned from these chains as burn-in and then they were combined (we used the species tree,

343 rather than the individual gene trees) into one large chain (of 7400M trees). Chains were

344 thinned by taking only each 5,000th tree. Using TreeAnnotator (from the BEAST 2 suite) we

345 constructed a Maximum Clade Credibility tree (using all 1.48M trees after thinning), storing

346 the mean heights.

347 We then pruned the tree from riverine species to obtain the pure Lamprologini tree on which

348 we fitted our model. Instead of performing one ABC-SMC inference on the obtained MCC

349 tree using a huge number of particles, which would be more accurate but computationally

350 extremely demanding, we performed 100 parallel inferences using 10,000 particles each. We

351 report the mean Bayes factor across these replicates.

352 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

353 Branching time uncertainty in the empirical tree

354 To account for uncertainty in the estimates of branching times in the Lamprologini tree we

355 sampled 100 trees from the posterior distribution obtained by *BEAST. Sampling was

356 performed at random, irrespective of the likelihood of the trees. In the Supplementary

357 material we show that the distribution of summary statistics of the subset of 100 trees is

358 similar to the distribution of the thinned chain. The 100 sampled trees were, like the

359 Maximum Clade Credibility tree, also pruned to remove the riverine taxa and stored

360 separately. For all 100 trees we performed both the ABC-SMC model selection algorithm and

361 the ABC-SMC parameter estimation algorithm, to determine the impact of different

362 branching times on the inferred water level model and associated parameters, and to

363 determine whether the MCC tree is a good representation of the underlying variability. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

364 RESULTS

365 Lamprologini phylogeny

366

367 The onset of diversification within the Lamprologini is estimated to be around 3.96 Ma (95%

368 Highest Posterior Density interval (HPD): [3.09, 4.91]), which is very close to the prior we

369 put on the node age, based on previous estimates (Meyer et al. 2016). Furthermore, we

370 estimate the branching off of the Congo species (N. congoensis and N. teugelsi) from the

371 Lamprologini in Lake Tanganyika to have occurred around 1.35 Ma (HPD: [1.12, 1.68]),

372 which is a bit younger than previously obtained estimates (1.70 Ma, (Sturmbauer et al.

373 2010)). The topology of the Maximum Clade Credibility tree is largely consistent with

374 previous findings (Sturmbauer et al. 2010) (Figure 1). Placement of Neolamprologus

375 fasciatus as a close relative to N. wauthioni seems to re-iterate previously published evidence

376 for introgressive hybridization (Koblmüller et al. 2007). For the three species not previously

377 included in the Lamprologini phylogeny, Lepidiolamprologus mimicus was placed as a close

378 relative to the other species within the genus Lepidiolamprologus, Chalinochromis

379 cyanophleps was placed as a sister species to Chalinochromis brichardi, within the group of

380 Chalinochromis and Julidochromis species, and in agreement with previous analysis

381 (Kullander et al. 2014b). In contrast to previous findings (Kullander et al. 2014b),

382 Neolamprologus timidus is not placed as a sister species to Neolamprologus furcifer, but

383 rather associates with N. mondabu and N. falcicula. Again, in contrast to other previous

384 findings (Gante et al. 2016), we place N. olivaceous outside the brichardi complex, which

385 includes the model system species N. brichardi and N. pulcher. (but see the DensiTree

386 representation, which shows that this is not true for all trees). Interestingly, we also do not

387 infer N. savoryi to be phylogenetically clustered within the brichardi complex (the ‘Princess

388 cichlids’ (Gante et al. 2016)), in contrast to Gante et al. (2016). We should take into account bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

389 however that the analysis by Gante et al. is based on on full genome sequences from only a

390 small group of species, in contrast to the limited number of markers from a large number of

391 species that we used.

392 As a reference we inferred speciation and extinction using the constant-rates birth-death

393 model (Nee et al. 1994). Using Maximum Likelihood (the function bd_ML in the DDD

394 package (Etienne et al. 2012)), we obtained an estimate of 1.871 myr-1 for the speciation rate,

395 and an estimate of 0.993 myr-1 for the extinction rate, for the MCC tree. We find an estimate

396 of 0.87 myr-1 for the diversification rate (speciation - extinction) and an estimate of 0.531

397 myr-1 for the turnover rate (extinction / speciation). For the 100 trees sampled from the full

398 chain, we obtain estimates of 3.02 myr-1 (95% HPD: [1.608, 4.947]) for the speciation rate

399 and 2.409 myr-1 (95% HPD: [0.884, 4.530]) for the extinction rate. This translates into

400 estimates of 0.61 myr-1 (95% HPD: [0.313, 0.985]) for the diversification rate, and 0.765 myr-

401 1 (95% HPD: [0.518, 0.930]) for the turnover rate. Estimates for the birth-death model

402 obtained during reconstruction of the tree using BEAST indicate an estimate of 0.864 myr-1

403 (95% HPD: [0.287, 1.459]) for the speciation rate and an estimate of 0.613 myr-1 (95% HPD:

404 [0.181, 0.953]) for the relative death rate, which translates into an estimate for the extinction

405 rate of 0.52 myr-1 (95% HPD: [0.156, 0.823]) per million years. This yields estimates of

406 0.334 myr-1 and 0.613 myr-1for the diversification and turnover rate respectively. The BEAST

407 inferences include the riverine species, so speciation and extinction rates are expected to be a

408 bit different. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

409

410 Figure 2. Phylogenetic hypothesis for the Lamprologini and outgroups, based on 3 411 mitochondrial and 6 nuclear genes, and two calibrations: 4 million years for the root of the 412 Lamprologini clade, and 1.1 - 3.5 million years for the Congo Lamprologini species. Left 413 panel: DensiTree (Bouckaert and Heled 2014) representation of the MCMC chain obtained 414 using *BEAST. Shown are trees from a thinned posterior chain, after selecting every 415 100,000th tree. Riverine species are indicated in grey. Right panel: Maximum Clade 416 Credibility tree. Bars around the node span the 95% HPD for each node. Please note that for 417 the dual display of both the densitree representation and the MCC phylogeny, some tips of 418 the MCC phylogeny might appear slightly misaligned. A high resolution version of both the 419 Densitree representation and the MCC phylogeny can be found in the supplementary 420 information. 421 422 Parameter estimation 423 We estimated parameter values for the three models for all 100 trees sampled from the

424 posterior. We report the parameter values across the combined posterior across all 100 trees.

425 Note that variation in the parameter estimates results from two sources of variation:

426 branching time variation across the 100 trees, and secondly variation in the parameter

427 estimate within each ABC-SMC inference.

428 The model without water level changes is identical to the constant-rates birth-death model,

# 429 and we find that our ABC-SMC estimates for sympatric speciation at high water level (. )

430 are slightly lower than the Maximum Likelihood estimate of the birth rate under the constant

431 rates birth-death model (2.644 Myr-1(95% HPD: [1.208, 4.633], see also Table 1) versus 3.02, bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

432 see also Table 1). Similarly, we infer the extinction rate (µ) to also be slightly lower (1.950

433 Myr-1 (95% HPD: [0.188, 4.101]) versus 2.409, see also Table 1). We obtain estimates of

434 0.694 and 0.738 for diversification and turnover respectively, which are close to the estimates

435 obtained using Maximum Likelihood (a diversification rate of 0.610 and a turnover rate of

436 0.765 respectively). Taking into account the 95% confidence intervals on the obtained

437 parameter estimates and the fact that the ABC-SMC estimates are potentially affected by the

438 prior while the ML estimates are not, we are confident that estimates obtained using our

439 ABC-SMC method for the model without water level changes are consistent with the

440 maximum likelihood estimates under the constant-rates birth-death model.

441 Using the LW model, which implements water level changes following the literature (e.g.

442 high water level until ~1.1 Ma, after which a series of water level changes took place), we

443 infer a lower rate of sympatric speciation at high water level (0.871 Myr-1 (95% HPD: [0.227,

444 3.642])), which is compensated with a high rate of allopatric speciation (6.412 Myr-1 (95%

445 HPD: [0.001, 14.195])) but not with a high rate of sympatric speciation at low water level

446 (0.028 Myr-1 (95% HPD: [0.001, 0.651])), suggesting that water level dynamics are important

447 drivers of biodiversity, but only through allopatric speciation. Extinction is inferred to be low

448 (0.037 Myr-1 (95% HPD: [0.001, 2.133])). Because of the non-trivial relationship between

449 speciation at high and low water level, we can no longer calculate diversification and

450 turnover rates.

451 Using the EW model, where water level changes are extrapolated beyond 1.1 Ma, we observe

452 that the rate of sympatric speciation at high water level is inferred to be similar to without

453 water level changes (2.753 Myr-1 (95% HPD: [1.347, 4.383])). Extinction, however, is lower

454 (0.111 Myr-1 (95% HPD: [0.001, 1.627])), and allopatric speciation and sympatric speciation

455 at low water level are both inferred to be much lower than for the literature water scenario bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

456 (0.022 Myr-1 (95% HPD: [0.001, 0.466]) and 0.033 Myr-1 (95% HPD: [0.001, 0.504])

457 respectively).

458 Across the three water level models we observe that the distribution of the post-hoc

459 perturbations does not differ substantially from the prior for the NW and EW water models,

460 with low estimates (0.036 (95% HPD: [0.001, 0.569]) and 0.030 (95% HPD: [0.001, 0.484])

461 for the NW and EW model respectively, Table 1). We notice a much higher value of

462 associated with LW (0.174, (95% HPD: [0.001, 0.680])), which also has a much higher

463 estimate for allopatric speciation at low water level. Allopatric speciation at low water level

464 potentially causes temporal alignment of branching times and we introduced the parameter

465 to correct simulated phylogenies for this, to allow comparison with phylogenies generated by

466 *BEAST, which does not allow for temporally aligned branching times. Hence, the higher

467 inferred value of for the LW model confirms the validity of the application of our post-hoc

468 perturbation.

469 470 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

# 471 Table 1. Median posterior density estimate, for sympatric speciation at high water (. ), ' 472 extinction (), perturbation ( ), sympatric speciation at low water (. ) and allopatric ' 473 speciation (). Shown are results for the model with no water level changes (NW), literature 474 values for water level changes (LW) and water level changes extrapolated beyond the 475 literature range (EW). The 95% credibility interval is shown between square brackets. All 476 values are rates per million years.

¾ Â Â õÉ ö ü õ· õÉ NW 2.644 [1.208, 4.633] 1.950 [0.188, 4.101] 0.036 [0.001, 0.569] LW 0.871 [0.227, 3.642] 0.037 [0.001, 2.133] 0.174 [0.001, 0.68] 6.412 [0.001, 14.195] 0.028 [0.001, 0.651] EW 2.753 [1.347, 4.383] 0.111 [0.001, 1.627] 0.030 [0.001, 0.484] 0.022 [0.001, 0.466] 0.033 [0.001, 0.504] 477

478 479 Figure 3. Posterior densities of the pooled posterior distribution across 100 randomly drawn 480 trees from the posterior MCMC chain. Shown are estimates for the three water level 481 scenarios (no water level changes (NW), literature values for water level changes (LW) and 482 extrapolated values for water level changes (EW)). Shown are the posterior density (black 483 line) and the 95% credibility interval (shaded area, blue for NW, gold for LW and green for 484 EW. X-axes are on a 10log scale. The first column shows a sample water level profile, with 485 the water level on the y-axis, and the time before present (in million years) on the x-axis. 486 Note that for the EW model, for each simulation a new profile was generated, and that the ' 487 shown profile is only one example of such a profile. Because without water level changes,  ' 488 and . have no meaning, their posterior distribution is not shown for the NW scenario. 489 490 491 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

492 Model fitting

493 494 Figure 4. Model selection results on 100 trees randomly drawn from the *BEAST posterior of 495 the Lamprologini tree. The top row shows 2 ln (Bayes factors) comparing posterior support 496 of the LW (literature water changes) model with the NW (no water level changes) model, the 497 bottom row shows 2 ln (Bayes factors) of the comparison between the posterior support for 498 the LW model with the EW (extrapolated water level changes) model. A 2ln(Bayes factor) 499 higher than 2 is generally considered to provide substantial evidence in favor of the 500 respective model (Kass and Raftery 1995), which is indicated by the solid lines. The thin 501 dotted line indicates the median 2 ln (Bayes factor) obtained for the MCC tree, for which we 502 do not find substantial support for any of the three models. The thick dotted line indicates the 503 median 2 ln (Bayes factor) for the trees drawn from the *BEAST posterior (e.g. the median of 504 the distribution shown), which is in both cases above 2, indicating substantial support for the 505 LW model compared to the other two models. 2 ln(Bayes factors) higher than 10 are grouped 506 together into one category. 507 508 509 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

510 Model selection 511 When we apply the model selection algorithm to the MCC tree, we find median Bayes factors

512 (we report here not the raw numbers, but 2 ln(Bayes factor), but for brevity refer to them as

513 Bayes factors) of 1.64 and 0.9 when comparing the LW model with the NW and EW model

514 respectively. We thus find no convincing evidence for any of the three models, when fitting

515 our model to the MCC tree. Alternatively, when we fit to 100 trees randomly sampled from

516 the *BEAST posterior, we find Bayes factors of 3.60 and 3.65 when comparing LW model

517 with the NW and EW model respectively. Furthermore, in 77 out of the 100 trees we select

518 the LW model as the most likely model (based on the Bayes factor), in 17 out of 100 trees we

519 select the EW model, and only in 6 out of 100 trees we select the model without any water

520 level changes.

521

522 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

523 Validation of the model selection procedure

524 525 Figure 5. Validation of the ability of our ABC-SMC algorithm to infer the correct model. 100 526 replicate datasets were generated for each water level model (no water level changes NW, 527 water level changes from the literature LW, or water level changes extrapolated beyond the 528 literature range, EW). The plots show the distribution of the 2 ln(Bayes factor) across all 100 529 replicate inferences. The dotted line indicates the median 2 ln(Bayes factor). A 2ln(Bayes 530 factor) higher than 2 is generally considered to provide substantial evidence in favor of the 531 respective model (Kass and Raftery 1995). 2 ln(Bayes factors) higher than 10 are grouped 532 together into one category. 533 534 Model validation shows that when we simulated data using the NW model, the NW model

535 was selected using our model validation algorithm more than the other two models (59 out of

536 100 replicates). Median Bayes factors are both higher than 2, with a median of 5.28 and 3.83

537 versus the LW and EW model respectively, supporting considerable support for the NW bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

538 model over the other two models. When data was simulated with the LW model, we selected

539 the correct model in the majority of 100 replicates (84 out of 100 replicates). The Bayes

540 factors reflect this, with medians of 18.4 (this is the maximum score) versus both the NW and

541 EW model. Lastly, when we simulated data using the EW model, we selected the correct

542 model more than the other two models, in 65 out of 100 replicates. This was reflected by the

543 Bayes factors as well, as the median Bayes factor versus the NW model was 18.4, and the

544 median Bayes factor versus the LW model was 7.47.

545 More interesting is the correct detection rate of a model, which is given by the number of

546 trees simulated by the model that is selected for that tree. This is equal to asking whether,

547 given posterior support for a respective model, we also find that the tree for which we find

548 this support was simulated with the respective model. If our model selection procedure can

549 not detect models accurately, we expect a detection rate of around 50%, as support is always

550 divided between two (not three) models. Detection rates larger than 50% support the

551 conclusion that our model selection procedure can adequately infer the correct model.

552 We find that across the 300 simulated trees, 120 trees received considerable support for the

553 LW model over the NW model (e.g. 2 ln (BF LW/NW) > 2), of these 120 trees, 92 trees were

554 simulated with the LW model, which leads to a correct detection rate of 77% (See Figure 6).

555 Furthermore, out of 107 trees that received considerable support for the LW model over the

556 EW model, we find that 83 trees were simulated using the LW model, which translates to a

557 detection rate of 78%. We find similar detection rates for the NW model: 90% against the

558 LW model (62 out of 69 detected trees) and 83% against the EW model (54 out of 65

559 detected trees). Lastly, detection rates for the EW model mirror these findings: a detection

560 rate of 79% against the NW model (79 out of 100 detected trees), and of 86% against the LW

561 model (68 out of 79 detected trees).

562 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

563

564 Figure 6. Accuracy of assignment of models depending on their posterior support. A: The 565 relative fraction of trees simulated with either LW (dark) or NW (light), receiving support for 566 LW (2ln(Bayes Factor LW/NW) > 2), support for NW (2ln(Bayes Factor LW/NW) < -2), or 567 receiving no support for either model. B: The relative fraction of trees simulated with either bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

568 LW (dark) or EW (light), receiving support for LW (2ln(Bayes Factor LW/EW) > 2), support 569 for EW (2ln(Bayes Factor LW/EW) < -2), or receiving no support for either model. C: The 570 relative fraction of trees simulated with either NW (dark) or EW (light), receiving support for 571 NW (2ln(Bayes Factor NW/EW) > 2), support for EW (2ln(Bayes Factor NW/EW) < -2), or 572 receiving no support for either model. 573 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

574 DISCUSSION

575 We have presented a model that infers past speciation and extinction rates, and their

576 interactions with changes in the environment, from a given phylogeny. We have shown that

577 our model is able to accurately select between different scenarios, including or excluding

578 environmental change. We applied our model to an updated phylogeny of the cichlid fish

579 tribe of Lamprologini and found evidence that past water level changes have shaped current

580 cichlid diversity in Lake Tanganyika, when we applied our model to a sample from the

581 posterior distribution of trees of the Lamprologini, as inferred by *BEAST. We asked the

582 model to select the best fitting of three scenarios: a scenario without any water level changes,

583 a scenario using the values found in the literature, and a scenario using the mean rate of water

584 level change found in the literature to extrapolate water level changes beyond the range of

585 literature values available. We found that the model following literature water levels received

586 most support, which suggests that water level changes have been an important driver of

587 diversity in the Lamprologini. We note that a model without effect of water level changes on

588 diversification (NW) can sometimes generate patterns that resemble the predictions of the

589 preferred model (LW). Yet, we find when fitting our model to trees drawn from the *BEAST

590 posterior that the distribution of Bayes Factors is skewed towards the model following

591 literature water levels (LW) and we find support for the model without an effect of water

592 level changes on diversification (NW) only for a small number of trees, suggesting that this

593 effect is relatively small.

594

595 When we applied our model selection algorithm on the Most Credible Consensus (MCC)

596 tree, we found contrasting results. Support for both models including water level changes

597 diminished, and posterior support for the model without any water level changes increased. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

598 Nevertheless, no single model could yield enough support to convincingly reject the other

599 two. Moreover, results using the MCC tree are markedly different from those using trees

600 sampled from the posterior. We conclude therefore that the MCC tree, at least for the

601 Lamprologini, but most likely more generally, provides a poor summary of the true species

602 tree and of the underlying variation in branching patterns. Hence, we suggest to avoid

603 reporting MCC trees, and instead to provide the reader with the full posterior distribution, for

604 instance through a DensiTree plot (Bouckaert and Heled 2014). Posterior inference, for

605 instance of speciation and extinction rates should preferentially also be performed on multiple

606 independent samples from the posterior, rather than on the MCC tree, as the underlying

607 variation might lead to very different results, as we have shown here.

608

609 Discrepancies between the MCC tree and the posterior distribution of trees could also

610 potentially clarify previously recovered inconsistencies when studying diversification, for

611 example in shrews in the Philippines. The Philippines have been subject to strong sea level

612 fluctuations, causing the fission and fusion of several islands, primarily during the

613 Pleistocene (Brown et al. 2013). Population genetic evidence has convincingly shown that the

614 location of such fused islands correlates strongly with genetic divergence between

615 populations in many different species (Evans et al. 2003; Linkem et al. 2010; Siler et al.

616 2010; Oaks et al. 2013). Phylogenetic analysis however, has failed to show any evidence of

617 diversification associated with Pleistocene water level changes (Esselstyn and Brown 2009).

618 The basis for this phylogenetic analysis however, was an MCC tree. Repeating the analysis

619 on the posterior distribution underlying the MCC tree could mitigate these problems, and

620 could clarify the impact of Pleistocene water level changes on diversification in the

621 Philippines archipelago. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

622

623 When allopatric speciation rates are high, the resulting phylogenetic trees have internal nodes

624 that have synchronized branching times, e.g. branching times that align with episodes of

625 water level change. Although Phylogenetic reconstruction software is able to infer

626 simultaneous branching events, it typically uses only two parameters (birth and death) to infer

627 all branching events of the tree. Therefore, if it can accommodate the simultaneous events, it

628 is unlikely to fit well to the non-simultaneous events, and vice-versa. Our finding of evidence

629 for a substantial role of habitat dynamics in diversification can therefore be regarded as

630 conservative. To improve the fit of trees generated by our model with trees generated by

631 *BEAST we included an a posteriori perturbation parameter in our model. This parameter

632 determines the standard deviation of a Gaussian perturbation kernel that is applied to each

633 node after the simulation has completed. By perturbing each node, we minimized the

634 probability that branching times align in time. We found that standard deviation increased in

635 size with an increase in allopatric speciation, as expected. A less ad hoc solution to deal with

636 the alignment of branching times in the tree would be to incorporate the model presented here

637 as a tree prior in phylogenetic reconstruction software. Although this need not introduce any

638 significant differences in the tree topology, the distribution of branching times could be

639 substantially influenced, and any subsequent inference focusing on such patterns could be

640 very different. Including such models in tree reconstruction software may require

641 incorporation of ABC methods, and will be extremely computationally demanding, but our

642 results justify such an endeavor.

643

644 Given that water level changes are only prevalent during the last million years before present,

645 we cannot exclude the possibility that increased diversification due to reasons other than bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

646 changing water levels has driven diversification during this period. On average, the LW

647 model could be represented by a simple birth-death model with a rate shift around one

648 million years ago. We expect however that although such a model could accommodate the

649 increased average diversification, it cannot replicate temporal alignment in branching events

650 due to water level changes. To examine this in more detail, we fitted a simple birth-death

651 model with a rate shift around one million years ago to the trees obtained from the posterior

652 (see Supplementary Information).. In the absence of a likelihood for the LW model, we

653 compared the nLTT statistic for the rate-shift model with that of the LW model, as the nLTT

654 statistic should be sensitive to detecting temporal alignment of branching events, We find that

655 our model is much closer to the empirical data than the rate shift model. We attempted to

656 improve the fit of the rate-shift model by allowing the speciation rate in the model to shift up

657 and down in line with the literature values of the water level changes. The two rates inferred

658 by the model then represent speciation at low, and at high water level respectively. Although

659 we do find an increase in the rate of speciation at low water level, the fit of this rate-shift

660 model is still worse than that of the LW model. This supports our conclusion that water level

661 changes influence the phylogeny not only through an increased speciation rate, but also

662 through temporal alignment of branching times.

663

664 Although we refer in our model to the different implementations of speciation as sympatric

665 and allopatric, care should be taken in interpreting these forms of speciation. We consider

666 here allopatric speciation only on a large scale, where populations become allopatric over

667 stretches of hundreds of kilometers (Sturmbauer et al. 2001). Large-scale isolation might not

668 be necessary for cichlids, as some species can already be limited in gene flow by a sand

669 stretch of 50 meters separating populations (Rico and Turner 2002). Such micro-allopatric

670 speciation events are not captured by the allopatric speciation rate in our model. Rather, these bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

671 local scale events are captured in our model by sympatric speciation. Hence, sympatric

672 speciation in our model covers all degrees of speciation ranging from full sympatry to

673 allopatry, providing that geographical isolation is smaller than that imposed by a water level

674 change. Allopatric speciation in our model then solely refers to speciation events caused by

675 geographical isolation over a large distance, driven by changes in water level, and inducing

676 simultaneous branching events.

677

678 In our model we have assumed that when the water level drops, species distribute themselves

679 equally over the two pockets of water that survive the water level drop. A more realistic

680 model would allow for a skew towards one of the pockets, either dependent on the respective

681 sizes of the pockets, the distribution of the species over the lake at high water level, or both.

682 We have here refrained from including a parameter that regulates the distribution of species

683 over the two pockets in order to avoid over fitting. Another possible extension of our model

684 would lie into extending the approach towards three or more pockets, possibly combined with

685 a parameter governing the distribution of species across these three pockets during a water

686 level drop. Bathymetric maps of Lake Tanganyika suggest that for some water level changes

687 it might split into three lakes (Coulter 1991). How a split of a species into three populations,

688 and associated allopatric divergence and speciation, affects phylogenetic structure and affects

689 temporal alignment in branching times remains currently unexplored and would be an

690 interesting avenue for future work.

691

692 Our results are strongly in line with population genomic analyses in a number of cichlid

693 species including Eretmodus cyanostictus (Verheyen et al. 1996), Tropheus moorii

694 (Koblmüller et al. 2011; Nevado et al. 2013; Sefc et al. 2017), Variabilichromis moorii bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

695 (Nevado et al. 2013), Altolamprologus (Koblmüller et al. 2016) and Telmatochromis

696 temporalis (Winkelmann et al. 2016), and resonate with population genomic findings across

697 the three African Rift Lakes (Sturmbauer et al. 2001). Furthermore, population genetic

698 studies have shown that water level fluctuations in Lake Malawi have been associated with

699 population expansion in cichlid species (Arnegard et al. 1999; Sturmbauer et al. 2001;

700 Genner et al. 2010), suggesting a potential role for water level changes in Lake Malawi as

701 well. Phylogenetic reconstruction for Malawi cichlid species is problematic however,

702 partially due to the young age of the species. However, considering that the geological record

703 of Lake Malawi spans a much larger part of the total lifespan of the lake (Delvaux 1995;

704 Lyons et al. 2015; Ivory et al. 2016) and thus provides a much better record about water level

705 fluctuations since the colonization of the lake by cichlids, we expect that modern genetic

706 developments will soon allow for a thorough understanding of the impact of water level

707 changes on cichlids in Lake Malawi as well.

708

709 Conclusion

710 Our model integrates standard constant-rate birth-death mechanics with environmental

711 change and with speciation induced by geographical isolation. We analyzed the phylogeny of

712 the tribe of Lamprologini to see whether past water level changes in Lake Tanganyika have

713 contributed to the current diversity of cichlid fish in Lake Tanganyika. We find an important

714 role for environmental changes in driving diversity, and find evidence that past water level

715 changes have shaped current standing diversity in the tribe of Lamprologini. However, we

716 found that inference of past environmental changes from a single phylogeny, and more

717 specifically, from the MCC tree, tends to lead to unreliable results. We therefore advocate

718 caution when using the MCC tree as a basis for further analysis. Furthermore, we argue for bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

719 the inclusion of more detailed branching models in phylogenetic reconstruction software,

720 which allow for the inclusion of an interaction between the environment, and speciation rates.

721

722 Acknowledgements

723 We thank Lucas Molleman for useful discussions. We thank the Netherlands Organisation for

724 Scientific Research (NWO) for financial support through VIDI and VICI grants awarded to

725 RSE. We thank the Donald Smits Center for information Technology of the University of

726 Groningen for their support and providing access to the Millipede and Peregrine high-

727 performance computing cluster. We thank the Max Planck Institute for Evolutionary Theory

728 for their support and providing access to their high-performance computing cluster. We thank

729 the Carl von Ossietzky Universität Oldenburg for their support and providing access to the

730 CARL computing cluster.

731 bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

732 REFERENCES

733 Aguilée R., Claessen D., Lambert A. 2013. Adaptive radiation driven by the interplay of eco- 734 evolutionary and landscape dynamics. Evolution (N. Y). 67:1291–1306. 735 Aguilée R., Lambert A., Claessen D. 2011. Ecological speciation in dynamic landscapes. J. 736 Evol. Biol. 24:2663–77. 737 Alin S., Cohen A., Bills R. 1999. Effects of landscape disturbance on animal communities in 738 Lake Tanganyika, East Africa. Conservation. 13:1017–1033. 739 Alin S., Cohen A.S. 2003. Lake-level history of Lake Tanganyika, East Africa, for the past 740 2500 years based on ostracode-inferred water-depth reconstruction. Palaeogeogr. 741 Palaeoclimatol. Palaeoecol. 199:31–49. 742 Arnegard M.E., Markert J.A., Danley P.D., Stauffer J.R., Ambali A.J., Kocher T.D. 1999. 743 Population structure and colour variation of the cichlid fishes Labeotropheus fuelleborni 744 Ahl along a recently formed archipelago of rocky habitat patches in southern Lake 745 Malawi. Proc. R. Soc. Lond. B. 266:119–130. 746 Barnosky A. 2005. Effects of Quaternary climatic change on speciation in mammals. J. 747 Mamm. Evol. 12:247–264. 748 Bouckaert R., Heled J. 2014. DensiTree 2: Seeing trees through the forest. bioRxiv.:1–11. 749 Bouckaert R., Heled J., Kühnert D., Vaughan T., Wu C.H., Xie D., Suchard M.A., Rambaut 750 A., Drummond A.J. 2014. BEAST 2: A Software Platform for Bayesian Evolutionary 751 Analysis. PLoS Comput. Biol. 10:1–6. 752 Brawand D., Wagner C.E., Li Y.I., Malinsky M., Keller I., Fan S., Simakov O., Ng A.Y., 753 Lim Z.W., Bezault E., Turner-Maier, J. Johnson J., Alcazar R., Noh H.J., Russell P., 754 Aken B., Alföldi J., Amemiya C., Azzouzi N., Baroiller J.-F., Barloy-Hubler F., Berlin 755 A., Bloomquist R., Carleton K.L., Conte M.A., D’Cotta H., Eshel O., Gaffney L., 756 Galibert F., Gante H.F., Gnerre S., Greuter L., Guyon R., Haddad N.S., Haerty W., 757 Harris R.M., Hofmann H. a., Hourlier T., Hulata G., Jaffe D.B., Lara M., A.P. L., 758 MacCallum I., Mwaiko S., Nikaido M., Nishihara H., Ozouf-Costaz C., Penman D.J., 759 Przybylski D., Rakotomanga M., Renn S.C.P., Ribeiro F.J., Ron M., Salzburger W., 760 Sanchez-Pulido L., Santos M.E., Searle S., Sharpe T., Swofford R., Tan F.J., Williams 761 L., Young S., Yin S., Okada N., Kocher T.D., Miska E. a., Lander E.S., Venkatesh B., 762 Fernald R.D., Meyer A., Ponting C.P., Streelman J.T., Lindblad-Toh K., Seehausen O., 763 Di Palma F. 2014. The genomic substrate for adaptive radiation in African cichlid fish. 764 Nature. 513:375–381. 765 Breman F.C., Loix S., Jordaens K., Snoeks J., Van Steenberge M. 2016. Testing the potential 766 of DNA barcoding in vertebrate radiations: the case of the littoral cichlids (Pisces, 767 Perciformes, Cichlidae) from Lake Tanganyika. Mol. Ecol. Resour. 16:1455–1464. 768 Brown R.M., Siler C.D., Oliveros C.H., Esselstyn J. a., Diesmos A.C., Hosner P. a., Linkem 769 C.W., Barley A.J., Oaks J.R., Sanguila M.B., Welton L.J., Blackburn D.C., Moyle R.G., 770 Townsend Peterson a., Alcala A.C. 2013. Evolutionary processes of diversification in a 771 model island archipelago. Annu. Rev. Ecol. Evol. Syst. 44:411–435. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

772 Capella-Gutiérrez S., Silla-Martínez J.M., Gabaldón T. 2009. trimAl: A tool for automated 773 alignment trimming in large-scale phylogenetic analyses. Bioinformatics. 25:1972– 774 1973. 775 Clabaut C., Salzburger W., Meyer A. 2005. Comparative phylogenetic analyses of the 776 adaptive radiation of Lake Tanganyika cichlid fish: nuclear sequences are less 777 homoplasious but also less informative than mitochondrial DNA. J. Mol. Evol. 61:666– 778 81. 779 Cohen A.S., Lezzar K.E., Tiercelin J.J., Soreghan M. 1997a. New palaeogeographic and lake- 780 level reconstructions of Lake Tanganyika: implications for tectonic, climatic and 781 biological evolution in a rift lake. Basin Res. 9:107–132. 782 Cohen A.S., Soreghan M., Scholz C.A. 1993. Estimating the age of formation of lakes: an 783 example from Lake Tanganyika, East African Rift system. Geology. 21:511. 784 Cohen A.S., Stone J.R., Beuning K.R.M., Park L.E., Reinthal P.N., Dettman D., Scholz C.A., 785 Johnson T.C., King J.W., Talbot M.R., Brown E.T., Ivory S.J. 2007. Ecological 786 consequences of early Late Pleistocene megadroughts in tropical Africa. Proc. Natl. 787 Acad. Sci. 104:16422–7. 788 Cohen A.S., Talbot M.R., Awramik S.M., Dettman D.L., Abell P. 1997b. Lake level and 789 paleoenvironmental history of Lake Tanganyika, Africa, as inferred from late Holocene 790 and modern stromatolites. Geol. Soc. Am. Bull. 109:444–460. 791 Coulter G. 1991. Lake Tanganyika and its life. Oxford University Press. 792 Coyne J., Orr H. 2004. Speciation. Sunderland, Massachusets U.S.A.: Sinauer Associates. 793 Day J.J., Santini S., Garcia-Moreno J. 2007. Phylogenetic relationships of the Lake 794 Tanganyika cichlid tribe Lamprologini: the story from mitochondrial DNA. Mol. 795 Phylogenet. Evol. 45:629–42. 796 Delvaux D. 1995. Age of Lake Malawi (Nyasa) and water level fluctuations. Mus. roy. Afrr. 797 centr., Tervuren (Belg.), Dept. Geol. Min., Rapp. ann. 1993 1994. 108:99–108. 798 Duftner N., Koblmüller S., Sturmbauer C. 2005. Evolutionary relationships of the 799 Limnochromini, a tribe of benthic deepwater cichlid fish endemic to Lake Tanganyika, 800 East Africa. J. Mol. Evol. 60:277–289. 801 Esselstyn J.A., Brown R.M. 2009. The role of repeated sea-level fluctuations in the 802 generation of shrew (Soricidae: Crocidura) diversity in the Philippine Archipelago. Mol. 803 Phylogenet. Evol. 53:171–181. 804 Etienne R.S., Haegeman B., Stadler T., Aze T., Pearson P.N., Purvis A., Phillimore A.B. 805 2012. Diversity-dependence brings molecular phylogenies closer to agreement with the 806 fossil record. Proc. R. Soc. B Biol. Sci. 279:1300–1309. 807 Etienne R.S., Rosindell J. 2012. Prolonging the past counteracts the pull of the present: 808 protracted speciation can explain observed slowdowns in diversification. Syst. Biol. 809 61:204–13. 810 Evans B.J., Brown R.M., McGuire J.A., Supriatna J., Andayani N., Diesmos A., Iskandar D., 811 Melnick D.J., Cannatella D.C. 2003. Phylogenetics of fanged frogs: testing bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

812 biogeographical hypotheses at the interface of the asian and Australian faunal zones. 813 Syst. Biol. 52:794–819. 814 Gante H.F., Matschiner M., Malmstrøm M., Jakobsen K.S., Jentoft S., Salzburger W. 2016. 815 Genomics of speciation and introgression in Princess cichlid fishes from Lake 816 Tanganyika. Mol. Ecol. 817 Genner M.J., Knight M.E., Haesler M.P., Turner G.F. 2010. Establishment and expansion of 818 Lake Malawi rock fish populations after a dramatic Late Pleistocene lake level rise. Mol. 819 Ecol. 19:170–82. 820 Genner M.J., Seehausen O., Lunt D.H., Joyce D.A., Shaw P.W., Carvalho G.R., Turner G.F. 821 2007. Age of cichlids: new dates for ancient lake fish radiations. Mol. Biol. Evol. 822 24:1269–82. 823 Glor R.E., Gifford M.E., Larson A., Losos J.B., Schettino L.R., Chamizo Lara A.R., Jackman 824 T.R. 2004. Partial island submergence and speciation in an adaptive radiation: a 825 multilocus analysis of the Cuban green anoles. Proc. R. Soc. B Biol. Sci. 271:2257–65. 826 Heaney L.R. 1985. Zoogeographic Evidence for Middle and Late Pleistocene Land Bridges 827 to the Philippine Islands. Mod. Quat. Res. Southeast Asia. 9:127–143. 828 Heled J., Drummond A.J. 2010. Bayesian Inference of Species Trees from Multilocus Data. 829 Mol. Biol. Evol. 27:570–580. 830 Hutter C.R., Guayasamin J.M., Wiens J.J. 2013. Explaining Andean megadiversity: The 831 evolutionary and ecological causes of glassfrog elevational richness patterns. Ecol. Lett. 832 16:1135–1144. 833 Ivory S.J., Blome M.W., King J.W., McGlue M.M., Cole J.E., Cohen A.S. 2016. 834 Environmental change explains cichlid adaptive radiation at Lake Malawi over the past 835 1.2 million years. Proc. Natl. Acad. Sci.:201611028. 836 Janzen T., Höhna S., Etienne R.S.R.S. 2015. Approximate Bayesian Computation of 837 diversification rates from molecular phylogenies: introducing a new efficient summary 838 statistic, the nLTT. Methods Ecol. Evol. 6:566–575. 839 Kass R., Raftery A. 1995. Bayes Factors. J. Amer. Stat. Assoc. 90:773–795. 840 Katoh K., Standley D.M. 2013. MAFFT multiple sequence alignment software version 7: 841 Improvements in performance and usability. Mol. Biol. Evol. 30:772–780. 842 Klett V., Meyer A. 2002. What, if anything, is a Tilapia?-mitochondrial ND2 phylogeny of 843 tilapiines and the evolution of parental care systems in the African cichlid fishes. Mol. 844 Biol. Evol. 19:865–883. 845 Koblmüller S., Duftner N., Sefc K.M., Aibara M., Stipacek M., Blanc M., Egger B., 846 Sturmbauer C. 2007. Reticulate phylogeny of gastropod-shell-breeding cichlids from 847 Lake Tanganyika--the result of repeated introgressive hybridization. BMC Evol. Biol. 848 7:7. 849 Koblmüller S., Nevado B., Makasa L., van Steenberge M., Vanhove M.P.M., Verheyen E., 850 Sturmbauer C., Sefc K.M. 2016. Phylogeny and phylogeography of Altolamprologus: 851 ancient introgression and recent divergence in a rock-dwelling Lake Tanganyika cichlid bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

852 genus. Hydrobiologia.:1–16. 853 Koblmüller S., Salzburger W., Obermüller B., Eigner E., Sturmbauer C., Sefc K.M. 2011. 854 Separated by sand, fused by dropping water: habitat barriers and fluctuating water levels 855 steer the evolution of rock-dwelling cichlid populations in Lake Tanganyika. Mol. Ecol. 856 20:2272–90. 857 Kocher T.D., Conroy J. a, McKaye K.R., Stauffer J.R., Lockwood S.F. 1995. Evolution of 858 NADH dehydrogenase subunit 2 in east African cichlid fish. Mol. Phylogenet. Evol. 859 4:420–432. 860 Kraaijeveld K., Kraaijeveld-Smit F.J.L., Maan M.E. 2011. Sexual selection and speciation: 861 the comparative evidence revisited. Biol. Rev. Camb. Philos. Soc. 86:367–77. 862 Kullander S.O., Karlsson M., Norén M. 2014a. Chalinochromis cyanophleps, a new species 863 of cichlid fish (Teleostei: Cichlidae) from Lake Tanganyika. Zootaxa. 3790:425–438. 864 Kullander S.O., Norén M., Karlsson M., Karlsson M. 2014b. Description of Neolamprologus 865 timidus, new species, and review of N . furcifer from Lake Tanganyika (Teleostei: 866 Cichlidae). Ichthyol.Explor.Freshwaters. 24:301–328. 867 Lanfear R., Calcott B., Ho S.Y.W., Guindon S. 2012. PartitionFinder: Combined selection of 868 partitioning schemes and substitution models for phylogenetic analyses. Mol. Biol. Evol. 869 29:1695–1701. 870 Lanfear R., Frandsen P.B., Wright A.M., Senfeld T., Calcott B. 2016. PartitionFinder 2: New 871 Methods for Selecting Partitioned Models of Evolution for Molecular and 872 Morphological Phylogenetic Analyses. Mol. Biol. Evol.:msw260. 873 Lezzar K.E., Tiercelin J.J., Batist M., Cohen A.S., Bandora T., Rensbergen P., Turdu C., 874 Mifundu W., Klerkx J. 1996. New seismic stratigraphy and Late Tertiary history of the 875 North Tanganyika Basin, East African Rift system, deduced from multichannel and 876 high-resolution reflection seismic data and piston core evidence. Basin Res. 8:1–28. 877 Linkem C.W., Hesed K.M., Diesmos A.C., Brown R.M. 2010. Species boundaries and 878 cryptic lineage diversity in a Philippine forest skink complex (Reptilia; Squamata; 879 Scincidae: Lygosominae). Mol. Phylogenet. Evol. 56:572–585. 880 Lyons R.P., Scholz C.A., Cohen A.S., King J.W., Brown E.T., Ivory S.J., Johnson T.C., 881 Deino A.L., Reinthal P.N., Mcglue M.M., Blome M.W. 2015. Continuous 1.3-million- 882 year record of East African hydroclimate, and implications for patterns of evolution and 883 biodiversity. PNAS.:2–7. 884 Matschiner M., Hanel R., Salzburger W. 2011. On the origin and trigger of the notothenioid 885 adaptive radiation. PLoS One. 6. 886 Matschiner M., Musilová Z., Barth J.M.I., Starostova Z., Salzburger W., Steel M., Bouckaert 887 R. 2016. Bayesian Phylogenetic Estimation of Clade Ages Supports Trans-Atlantic 888 Dispersal of Cichlid Fishes. Syst. Biol. 889 McGlue M.M., Lezzar K.E., Cohen A.S., Russell J.M., Tiercelin J.-J., Felton A. a., Mbede E., 890 Nkotagu H.H. 2008. Seismic records of late Pleistocene aridity in Lake Tanganyika, 891 tropical East Africa. J. Paleolimnol. 40:635–653. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

892 Meyer B.S., Matschiner M., Salzburger W. 2015. A tribal level phylogeny of Lake 893 Tanganyika cichlid fishes based on a genomic multi-marker approach. Mol. Phylogenet. 894 Evol. 83:56–71. 895 Meyer B.S., Matschiner M., Salzburger W. 2016. Disentangling incomplete lineage sorting 896 and introgression to refine species-tree estimates for Lake Tanganyika cichlid fishes. 897 Syst. Biol. syw069:1–20. 898 Muschick M., Indermaur A., Salzburger W. 2012. Convergent Evolution within an Adaptive 899 Radiation of Cichlid Fishes. Curr. Biol.:1–7. 900 Nagai H., Terai Y., Sugawara T., Imai H., Nishihara H., Hori M., Okada N. 2011. Reverse 901 evolution in RH1 for adaptation of cichlids to water depth in Lake Tanganyika. Mol. 902 Biol. Evol. 28:1769–1776. 903 Nee S., May R.M., Harvey P.H. 1994. The reconstructed evolutionary process. Philos. Trans. 904 R. Soc. Lond. B. Biol. Sci. 344:305–11. 905 Nevado B., KoblmÜller S., Sturmbauer C., Snoeks J., Usano-Alemany J., Verheyen E. 2009. 906 Complete mitochondrial DNA replacement in a Lake Tanganyika cichlid fish. Mol. 907 Ecol. 18:4240–4255. 908 Nevado B., Mautner S., Sturmbauer C., Verheyen E. 2013. Water-level fluctuations and 909 metapopulation dynamics as drivers of genetic diversity in populations of three 910 Tanganyikan cichlid fish species. Mol. Ecol. 22:3933–48. 911 O’Quin K.E., Hofmann C.M., Hofmann H.A., Carleton K.L. 2010. Parallel Evolution of 912 opsin gene expression in African cichlid fishes. Mol. Biol. Evol. 27:2839–2854. 913 Oaks J.R., Sukumaran J., Esselstyn J. a, Linkem C.W., Siler C.D., Holder M.T., Brown R.M. 914 2013. Evidence for climate-driven diversification? A caution for interpreting ABC 915 inferences of simultaneous historical events. Evolution (N. Y). 67:991–1010. 916 Papadopoulou A., Knowles L.L. 2015. Genomic tests of the species-pump hypothesis: Recent 917 island connectivity cycles drive population divergence but not speciation in Caribbean 918 crickets across the Virgin Islands. Evolution (N. Y). 69:1501–1517. 919 Pearse W.D., Purvis A. 2013. phyloGenerator: An automated phylogeny generation tool for 920 ecologists. Methods Ecol. Evol. 4:692–698. 921 Pybus O., Harvey P. 2000. Testing macro–evolutionary models using incomplete molecular 922 phylogenies. Proc. R. Soc. B Biol. Sci. 267:2267–72. 923 Rico C., Turner G.F. 2002. Extreme microallopatric divergence in a cichlid species from 924 Lake Malawi. Mol. Ecol. 11:1585–90. 925 Rossiter A. 1995. The cichlid fish assemblages of Lake Tanganyika: ecology, behaviour and 926 evolution of its species flocks. Adv. Ecol. Res. 927 Salzburger W., Baric S., Sturmbauer C. 2002. Speciation via introgressive hybridization in 928 East African cichlids? Mol. Ecol. 11:619–25. 929 Santos M.E., Braasch I., Boileau N., Meyer B.S., Sauteur L., Böhne A., Belting H.-G., 930 Affolter M., Salzburger W. 2014. The evolution of cichlid fish egg-spots is linked with a bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

931 cis-regulatory change. Nat. Commun. 5:5149. 932 Schelly R., Salzburger W., Koblmüller S., Duftner N., Sturmbauer C. 2006. Phylogenetic 933 relationships of the lamprologine cichlid genus Lepidiolamprologus (Teleostei: 934 Perciformes) based on mitochondrial and nuclear sequences, suggesting introgressive 935 hybridization. Mol. Phylogenet. Evol. 38:426–438. 936 Schelly R., Takahashi T., Bills R., Hori M. 2007. The first case of aggressive mimicry among 937 lamprologines in a new species of Lepidiolamprologus (Perciformes: Cichlidae) from 938 Lake Tanganyika. Zootaxa. 49:39–49. 939 Schwarzer J., Misof B., Tautz D., Schliewen U.K. 2009. The root of the East African cichlid 940 radiations. BMC Evol. Biol. 9:186. 941 Schweiger O., Klotz S., Durka W., Kühn I. 2008. A comparative test of phylogenetic 942 diversity indices. Oecologia. 157:485–95. 943 Sedano R.E., Burns K.J. 2010. Are the Northern Andes a species pump for Neotropical birds? 944 Phylogenetics and biogeography of a clade of Neotropical tanagers (Aves: Thraupini). J. 945 Biogeogr. 37:325–343. 946 Seehausen O. 2000. Explosive speciation rates and unusual species richness in 947 haplochromine cichlid fishes: effects of sexual selection. Adv. Ecol. Res. 31:237–274. 948 Seehausen O. 2006. African cichlid fish: a model system in adaptive radiation research. Proc. 949 R. Soc. B Biol. Sci. 273:1987–98. 950 Sefc K.M., Mattersdorfer K., Ziegelbecker A., Neuhüttler N., Steiner O., Goessler W., 951 Koblmüller S. 2017. Shifting barriers and phenotypic diversification by hybridisation. 952 Ecol. Lett. 953 Shirai K., Inomata N., Mizoiri S., Aibara M., Terai Y., Okada N., Tachida H. 2014. High 954 prevalence of non-synonymous substitutions in mtDNA of cichlid fishes from Lake 955 Victoria. Gene. 552:239–245. 956 Siler C.D., Oaks J.R., Esselstyn J.A., Diesmos A.C., Brown R.M. 2010. Phylogeny and 957 biogeography of Philippine bent-toed geckos (Gekkonidae: Cyrtodactylus) contradict a 958 prevailing model of Pleistocene diversification. Mol. Phylogenet. Evol. 55:699–710. 959 Spady T.C., Seehausen O., Loew E.R., Jordan R.C., Kocher T.D., Carleton K.L. 2005. 960 Adaptive molecular evolution in the opsin genes of rapidly speciating cichlid species. 961 Mol. Biol. Evol. 22:1412–1422. 962 Sparks J.S., Smith W.L.W.L. 2004. Phylogeny and biogeography of cichlid fishes (Teleostei: 963 Perciformes: Cichlidae). Cladistics. 20:501–517. 964 Sturmbauer C., Baric S., Salzburger W., Rüber L., Verheyen E. 2001. Lake level fluctuations 965 synchronize genetic divergences of cichlid fishes in African lakes. Mol. Biol. Evol. 966 18:144–54. 967 Sturmbauer C., Salzburger W., Duftner N., Schelly R., Koblmüller S. 2010. Evolutionary 968 history of the Lake Tanganyika cichlid tribe Lamprologini (Teleostei: Perciformes) 969 derived from mitochondrial and nuclear DNA data. Mol. Phylogenet. Evol. 57:266–84. bioRxiv preprint doi: https://doi.org/10.1101/085431; this version posted December 22, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

970 Sugawara T., Terai Y., Okada N. 2002. Natural Selection of the Rhodopsin Gene During the 971 Adaptive Radiation of East African Great Lakes Cichlid Fishes. Mol. Biol. Evol. 972 19:1807–1811. 973 Thorpe R.S., Surget-Groba Y., Johansson H. 2008. The relative importance of ecology and 974 geographic isolation for speciation in anoles. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 975 363:3071–81. 976 Toni T., Stumpf M.P.H. 2010. Simulation-based model selection for dynamical systems in 977 systems and population biology. Bioinformatics. 26:104–110. 978 Toni T., Welch D., Strelkowa N., Ipsen A., Stumpf M.P.H. 2009. Approximate Bayesian 979 computation scheme for parameter inference and model selection in dynamical systems. 980 J. R. Soc. Interface. 6:187–202. 981 Turner G.F., Seehausen O., Knight M.E., Allender C.J., Robinson R. 2001. How many 982 species of cichlid fishes are there in African lakes? Mol. Ecol. 10:793–806. 983 Vaidya G., Lohman D.J., Meier R. 2011. SequenceMatrix: Concatenation software for the 984 fast assembly of multi-gene datasets with character set and codon information. 985 Cladistics. 27:171–180. 986 Verheyen E., Rüber L., Snoeks J., Meyer A. 1996. Mitochondrial phylogeography of rock- 987 dwelling cichlid fishes reveals evolutionary influence of historical lake level fluctuations 988 of Lake Tanganyika, Africa. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 351:797–805. 989 Wagner C.E., Harmon L.J., Seehausen O. 2012. Ecological opportunity and sexual selection 990 together predict adaptive radiation. Nature. 487:366–9. 991 Wagner C.E., Harmon L.J., Seehausen O. 2014. Cichlid species-area relationships are shaped 992 by adaptive radiations that scale with area. Ecol. Lett. 17:583–592. 993 Wagner C.E., McIntyre P.B., Buels K.S., Gilbert D.M., Michel E. 2009. Diet predicts 994 intestine length in Lake Tanganyika’s cichlid fishes. Funct. Ecol. 23:1122–1131. 995 Weir J.T. 2006. Divergent Timing and Patterns of Species Accumulation in Lowland and 996 Highland Neotropical Birds. Evolution (N. Y). 60:842–855. 997 Weiss J.D., Cotterill F.P.D., Schliewen U.K. 2015. Lake Tanganyika — A ’ Melting Pot ’ of 998 Ancient and Young Cichlid Lineages ( Teleostei: Cichlidae )? :1–29. 999 Winkelmann K., Rüber L., Genner M.J. 2016. Lake level fluctuations and divergence of 1000 cichlid fish ecomorphs in Lake Tanganyika. Hydrobiologia.:1–14.

1001

1002